Field of the Invention
The invention relates to a method and an apparatus for processing audio signal, and more particularly, to a method and an apparatus for signal extraction of audio signal.
Description of Related Art
Generally, during a processing procedure of an audio signal such as voice or music, an ideal signal is maintained in the audio signal and a noise is removed from the audio signal. Ideal signal and noise segmentation may include a noise detection method and a signal extraction method. The noise detection method includes the following methods: an energy detection method using amplitude, power spectral density (PSD), zero crossing rate (ZCR) or the like; a model comparison method using Probability Model, Spectrum Model, Likelihood or the like; an auto convergence method using least mean square (LMS), normalized least mean square (NLMS) or the like; and an adaptability estimation method using Adaptive Filter, Moving Average, linear predictive coding (LPC) or the like.
Among them, the energy detection method and the model comparison method usually distinguishes the ideal signal from the noise on the time axis. The auto convergence method is incapable of separating frequency bands of the ideal signal and the noise for further analysis. As for the adaptability estimation method, the estimation may be inaccurate when the signal-to-noise ratio (SNR) is low.
In addition, the methods using signal extraction (including spectrogram 2D masking, signal model comparison, etc.) mostly belong to determination and identification for the known signal types. Those methods can only extract the expected signal types and may consume a lot of resources if there are too many signal types.